SRC TR 87-139 Digital Controllers for SISO Systems: A Review and a New Algorithm by E. Zafiriou and M. MorariINT. J. CONTROL, 1985, VoL. 42, NO. 4, 855-876 al controllers for SISO systems: a review and a new algorithm EVANGHELOS ZAFIRIOUt and MANFRED MORARIt Several digital control algorithms for linear single-input single-output systems are examined and the effect ofthe sampling period on their performance is analysed in terms of rippling. overshoot and settling time. The problem is addressed in the frequency domain (:-transform) and itis shown that each controller works for some classes of systems but that none works for all, The similarities and diflerences of these controllers are established and an explanation of their deficiencies is given based on the location of the zeros of the discrete system. The insight gained leads to. simple new rule for the design ofa controller which combines the advantages of the diferent algorithms but at the same time is free of their problems. A single tuning parameter is included which directly affects the closed-loop speed of response and bandwidth, The parameter can be used to detune the controller in the event that the real system dilfers from the model on which the controller design is based. No tuning is necessary when the available mode! is exact, unless smaller Values for the manipulated variable, atthe cost of a slower response, are preferred 1. Introduction In the literature one can find a very large number of digital control algorithms for single-loop systems, Numerous design criteria are used and often they are form- ulated so as to satisfy the special requirements ofa specific process. The synthesis of ‘many of those controllers is quite complicated and it is usually very difficult to know when and why problems may occur. ‘A common property of the control algorithms examined in this paper is the computational simplicity of their synthesis: it so happens that all of them can in fact be described as pole-rero placement controllers, though that may not be the way they were initially designed. Each of them tries to satisfy some commonly accepted criteria, such as: (i) A performance eriterion for the discrete output of the system (for example, minimization of the sum of squared errors) (ii) Behaviour between the sampling points: hidden oscillations in the system output can be caused by ringing of the manipulated variable or by unob- servable oscillatory open-loop modes. (ii) Setting time (for example, the requirement to reach the set-point in a finite number of time steps) (iv) Overshoot andor undershoot. (v) Manipulated variable: large values must be avoided because they usually cannot be implemented due to saturation of the manipulated variable. In this paper we compare the set-point step responses obtained with various controllers, In the first part, an analysis of the problems is given for the case of no Received 20 November 1984 + Department of Chemical Engincering, California Institute of Technology, Pasadena, California 91125, USA.856 E. Zafiriow and M. Morari modelling error and a new algorithm which avoids the disadvantages of the examined controllers is proposed. The second part deals with stability issues of the new algorithm when 3 mismatch between the model and the plant exists. 2. Evaluation of the control algorithms Inthe case of no modelling eror, the classical feedback (Fig. 1 (a) is equivalent, for command following, to the open-loop structure shown in Fig, 1 (b), where G.(=) may be considered as the controller to be designed There isa simple relation between G,(2)and the controller ofthe classical feedback structure Cz): = coe 1+ Cz)G(z) where G(z) isthe transfer function of the plant. We can similarly obtain C(z) from G.a) by a G6.) as Q) i) i Designing G.(2) and then obtaining C(z) by eqn. (2) is an established technique (see, for example, Jury and Schroeder (1956), This technique will be used through- out the paper because it allows presentation of the results in a transparent manner. For the pulse transfer function of the plant we shall use the expression 6) where 1V is the largest integer such that NT is less than or equal to the dead-time (T being the sampling time). The superscripts *—" and "+" denote the zeros inside and outside the unit circle respectively and k is the number of zeros inside the unit circle. G(z) is assumed to be open-loop stable. ep oe} fa @ : G ¢ Le ) Lr a}—Ls FIUTER CONTROLLER PLANT o Figure 1. (a) Classical feedback structure. (b) Open-loop structure equivalent to the feed back for the exact model. (c) The open-loop structure with the filter included.Digital controllers for SISO systems 857 2.1. Digital control algorithms (i) Dahtin’s controller This is a well known controller included in current textbooks like Smith (1972) and. Kuo (1977). It was introduced by Dahlin (1968), who proposed a method for designing the controller by specifying the closed-loop transfer function V(2) to be first order with a dead-time equal to that of the plant and a steady-state gain of 1-0, thus: [-exp(-T/a}-™ exp (= T/A 4 where 2 is the time constant of the closed-loop response and T is the sampling interval. From V(2) and G(:), the classical feedback controller is found to be (exp (— Tay)" 1 Cocle) exp (= T/A The time constant of the closed-loop response serves as a tuning parameter for adjusting the speed of the response. If we take the equivalent structure of Fig. 1 (b) the controller is found from eqn. (1) tobe 6 a) 6 =exp(=Tiae | K where the expression for G(z) given by (3) is used. Hence we can rearrange to get the structure of Fig. I {c) with controller Gele)= a and a filter t—exp(= T/A) exp (— T/A 6) ‘The reason for separating out F(z) and calling it a ‘iter’ will become apparent in § 40f the paper dealing with modellingerrors. Note that for i = 0(no filter), Dahlin's controller becomes a deadbeat controller which brings the output to the set-point value after one sampling period. The stability problems are discussed in § 22. Dahlin also proposed & modification to be used when there are problems with the inital algorithm; we will examine this modified controller in detail in §2.2, ater the reasons for the problems are explained. (Gi) The controller that minimizes the sum of the squared errors of the output This controller is designed so as to minimize the objective function = Fu Be where y; is the value of the system output at £ ‘T and r the desired set-point. The858 E. Zafiriow and M. Morari problem is solved in the z-domain by Chang (1961) and in the state space of Kucera (1972), ‘The controller for the structure of Fig. 1 (b) is (Kucera 1972) 2-45) Gi) Ouxput and state deadbeat controllers ‘The eurput deadbeat controller is defined as the controller that drives the discrete ‘output of the system to the set-point value and keeps it there, after a minimum number of time steps. The earliest discussion of this kind of controller was given by Bergen and Ragazzini (1954). Kalman (1954) raised the question of the behaviour of the continuous output between the sampling points. Jury and Schroeder (1956) and Jury (1958) studied the problem by using the modified z-transform and Kalman and Bertram (1959) in the state-space ‘The design of the state deadbeat controlier is based on the idea of having the states of the system at equilibrium when the discrete output reaches the set-point so that the continuous output will remain there as well. The minimum number of time steps required for all the states to be at equilibrium is at most equal to the number of the states, which is equal to the order of the system. For SISO systems it is easier to exp (—T/2)= 1/1223) Vogel and Edgar did not provide any guidelines for the selection of They suggest gradually increasing 2 until the response becomes satisfactory. In the case where there are two or more zeros near +1, has to be made very large in order to bring the overshoot or undershoot down to acceptable values (cancellation ‘with only one of the zeros is not sufficient). The result is that the response becomes to0 slow to be acceptable, This will happen for systems with two or more zeros in the Laplace transfer function, either in the LHP or the RHP. The problem is illustrated in Example 7. Example 7 E (s+ 16642) (e+ OSs 1S)(eF ISS +6) (2-005) 08192 +0951) (2— 0551) — 0861)? — 17262 + 0-779)Digital controllers for SISO systems 869 Two of the zeros are close enough to +1 to produce the extremely large overshoot shown in Fig. 11 (for 4= 0). By selecting 2 such that exp (— 7/2) is equal to one of the zeros (2=05, exp(~T/2)=0819, the response shown in Fig. 11 is obtained. For exp(—T/) equal to the other zero (2.= 1,exp(-~ T/A) =0905), the response is given in Fig. 12. The overshoot is still clearly unacceptable. In order to reduce it to 75%, 2 has to be increased to 2S and extremely sluggish behaviour results, (Fig 12), 1 1 o I 2 3 0 I 2 3 rime Figure 1. Gi) = 22854 Ns 2116+ 05) 15) 425546, T= Ole ~ ste deadbeat {ah Vogel Edgar 0, —~ Voge Edgar, 208, serpin $00 | This id 500 ; ; ° I 2 3 5 i\ Easdit TINE Figure 12. Gls) = 2.2564 Nis-+ Ils 4 O5)ls + F5Yls? + 255 4 6), T- ‘ie ty Vogel-Edgar, 2=2'5, —— the new algorithm, 2 =0; Vogel-Edgar, “set-point,870 E. Zafiriow and M. Morari 23. Discussion There are two basic problems associated with the performance of the controllers examined. One is rippling of the system output between the samples due to ringing ‘of the manipulated variable and the other is large overshoot and/or undershoot of the discrete output itself, Dablin’s controller will also be unstable when the system has zeros outside the unit circle and the modified Dahlin controller has to be stabilized in this case by appropriate selection of the tuning parameter. The transfer functions of the controller G,(2) of the open-loop structure (Fig. 1(b)) are given in Table 1. In §2.2 we explained how the problems are related to the location of the zeros of the model and how the location depends on the continuous system Laplace transfer function and on the sampling period T (Theorems 1 and 2). The relations between the continuous system, the discrete controller, the sampling period and the associated problems are given in Table 3. The entries in the first Continuous Controt system algorithm Small T Ty, -AT T,+aT pe = Unstabiet op Overshoott LHP zero SE : = “ sD Overshoot Overshoott —_Overshoott VE One zero: — More: overshoot Overshoott_——_Overshoott pe Unstable Unstable Rippling op Undershoot Rippling RHP zero SE = Rippling Rippling sD Undershoot “ ie VE One zero: = = More: undershoot pe n—m=2 rippling n—m>4 unstable Unstable Rippling Excess zer0s. = OD. Rippling = Rippling n= mevens SE Rippling Rippling Rippling sD ~ ~ = VE - = pe Unstable Unstable Rippling Excess zeros. OD. = = Ri n— mod SE = Rippling Rippling ee) sD = = _ VE = = = {Two or more zeros (sce §2.3) For n-m=2 only the column for small Tapphies DC Dahlin’ controller OD Output deadbeat controller ‘SE Minimizes sum of squared errors of the output 'SD_ State deadbeat controller VE Voge!-Edgar controler T_ Sampling period at which the ato ofthe dacrete system crvses the uni isle AT Range atound T, for which the corresponding problem eecure ‘mn Degrees of merator and denominator of Gs) ‘Table 3. Summary, of the problems of the examined control algorithms.Digital controllers for SISO systems 871 column refer to the Laplace transfer function of the system; hence the table shows what effect a LHP or a RHP zero or a pole excess (n ~ m) of 2. or more, can have on the performance of each controller for various ranges of the sampling period T: How small Thas to be in order for the problems in Table 3 to arise, depends on the continuous system and the type of problem. An overshoot of 100% caused by a LHP zero at small Ts can occur for T'< 0-75: for areal zero and for T< 09: for a pair of complex-conjugate zeros, where ris the inverse of the distance of the zero from the origin. Similarly for a RHP zero, we may get a significant undershoot for T<075¢ for a real zero and for T<09: for complex zeros. The rippling caused by excess zeros at small Ts (when (n~ m) is even) can occur for T up to 075tp, where tp is the dominant time constant of the system, but usually the problem appears for smaller Ts which can start as low as 0-2tp, To, denotes the critical sampling period at which @ zero crosses the unit circle. Rippling may occur around T,, for a range from 7.,—AT to T,,+4T. The value of T,, for each zero of G{2) has to be calculated from the Laplace transfer function G(s). AT appears to be a fraction of T,: it is about 037,, when we have a zer0 of Giz} corresponding to a RHP zero of G(s) and about 0-77,, when we have an excess zero crossing the unit circle, The reason for the difference is that a zero corresponding to a RHP zero covers a greater distance on the complex plane than an excess zero, as T goes from zero to infinity, and therefore it moves faster with T When G(s) has more than one zero in the LHP and when these zeros are much smaller than the poles (by an order of magnitude or more), then the corresponding discretized system G(:) can have a zero either inside or outside the unit circle, ending on the sampling time. From Theorems 1 and 2, it follows that if such a zero crosses the unit circle, then there are two T,s for which a crossing occurs. For the smaller T., at which the zero leaves the unit circle, the problems of the examined algorithms are listed in Table 3. For the larger T,, (at which the zero returns to the unit circle), all algorithms will generate significant overshoot before t= T and, in addition, have all the other problems particular {0 each algorithm. The reason is that the open-loop step response of systems of this type shows large overshoot and the second T,, is larger than the time 1 at which this overshoot occurs. This problem is not of practical importance since such a large T would result in a folding frequency x/T-smaller than the bandwidth of Gis) and would therefore not be selected. From an inspection of the open-loop step response, it can easily be determined if any of the zeros of G(z) arising from LHP zeros of Gis) cross the unit circle as follows. Assume that G(s) has been scaled so that G(0)>0. Ifthe unit step response is negative for some t’ and if T=" then G(z) has an odd number of zeros on the real axis between +] and +00, The converse is also true, The reason is that K in (3)is the value ofthe open-loop step response at '= T. For all the systems tested with such crossings there was always a range of Tfor which the number of zeros between +1 and +2 was odd. Hence, the above condition on the step response seems to be both necessary and sufficient for the existence of such crossings. However no proof is available 3. A new algorithm In §2 it became clear that the controller G,(2) of the structure of Fig. 1 (b) should not have poles with negative real part close to the unit circle, These poles cause oscillation of the controller output (with period equal to twice the sampling period if872 E. Zafiriow and M. Morari the poles are negative reals) and as a result the process output exhibits intersample rippling. Hence, when there are zeros of Gi) with negative real part, one should not use those zeros, or their inverses, as poles of G,(2), but use poles at the origin instead. When there are unstable zeros of G(z) with positive real part, one should use their inverses as poles of G.(2). If zeros close to +1 were substituted with poles at the origin, then, in order to avoid steady-state offset (ie. in order to have G.(1) = G(1)"), large undershoot or overshoot would have to be accepted. In addition, for all the control algorithms examined in § 2, the zeros of G.(=) ate chosen to be equal to the poles of Gtz). Hence a simple rule to design the controller G,(2) is the following. (1) Use as zeros of G., the poles of G(z) (they are assumed to be stable. 2) Use as poles of G,, the zeros of G(z) with positive real part which are inside the unit circle the inverses of those with positive real part which are outside the unit circle and as many at the origin as there are zeros with negative real part. (3) An additional pole of G, at the origin must be present Because of the inherent time delay of a discrete system. (4) The steady-state gain of G, should be: 1 C= ay Let the superscripts ‘—" and *4" denote ‘inside’ and ‘outside’ the unit circle, respectively, and the subscripts ‘+" and ‘positive’ and ‘negative’ real part. We have (1 1 aa ) ( 3) = Pad en ara(e- where k 2e10s (23,4, 03,4) have positive real part and are inside the unit circle, TR(@2 24,0442) have positive real part and are outside the unit circle and N= 1=TQ 1445s 2g) have negative real part A filter of the form Fle) =(1~a)(l—a2~') (where 0<2.<1) should also be included. Theoretical justification for the inclusion of the filter will be given in §4. At this point one can, nevertheless, note that the tuning parameter can be used to lower the values ofthe manipulated variable. This of course will be accomplished at the cost of a slower response. ‘The classical feedback controller C(z)for the structure of Fig. 1 (a) can be obtained from F)G.2) FOG.) ce) en i) ‘The behaviour of the new algorithm was examined for the examples of §2 and, as the corresponding plats show, it performs very well (Figs. 3-12)Digital controllers for SISO systems 873 ‘The algorithm is expected to work well all the time, unless of course there are ‘unobservable open-loop oscillatory modes (Astrém and Wittenmark 1984). This will happen if the sampling period matches the frequency of any oscillatory mode of the continuous open-loop system. ‘The problem can be met by appropriate selection of the sampling period (for example, less than half the period of oscillations of the continuous system impulse response (Jury 1957)). The controller G,(2) can then be designed from G{2) as above It should also be noted that this design procedure can be extended to external system inputs other than steps. One might already have observed that one can obtain the new algorithm of eqn. (26) by writing the controller transfer function Gse(2) (structure of Fig. 1 (b) that minimizes the sum of the squared errors and substituting the poles with negative real part with poles at the origin, while kecping the same steady-state gain. The same procedure can be used for inputs other than steps to yield a problem-free controller. One should first obtain the transfer function of the controller (structure of Fig. 1 (b) that minimizes the sum of the squared errors for the specified input and then make the correction mentioned above. 4. Sta For no modelling error, the classical feedback structure of Fig. 1 (a) is equivalent to the open-loop structure of Fig. 1 (b, where eqns. (1) and (2) relate the controllers C(z) and G(s). In the case where the model is not exact, et us denote the model by tz) and the actual plant by Giz). Then the feedback structure is equivalent to that of Fig. 13 (a) where the two added G2) blocks cancel each other. This structure can be simplified to that in Fig. 13 (b) The relationships between the controller G,(2) in Fig. 13(6) and the classical feedback controller C(z) are ty in the presence of modelling error 28) @9) For an exact model, the structure of Fig. 13 (b) simplifies for command-following to that shown in Fig. 1 (b), ie. it becomes open-loop, which makes the design of G. transparent. For example, if Dahlin had substituted the zeros which cause ringing in the corresponding G,() instead of C(z) he would have got the desired result regarding ringing, without the complications and problems of the modified Dahlin controller. ‘On the other hand, when the model Giz) is not exact, the structure of Fig, 13 () is not effectively open-loop and stability problems can arise. A very useful property of this structure is that any controller G2) which satisfies G.0) = G1)" yields zer0 offset (Garcia and Morari 1982). From Fig. 13 (b), we find (30) __ GIG.) en FG@lGe—cay" where wis the manipulated variable or stability, ts necessary and sufficient that both of the following characteristic equations have their roots strictly inside the unit crcl:874 E. Zafiriou and M. Morari & c conTagrces | PLANT i Y F & 6 FIcTeR convaocten| PUNT ee LL os Figure 13. (a)Structure equivalent to the classical feedback. (b) Simplification of ‘the Structure shown in (a) (€} The equivalent structure with the filter included, Gp tte) Gen=0 (32) + ie(2) By adding @ filter Fle) to the controller G,(z) we get the structure in Fig 13(c). The importance of the filter is shown by the following theorem (Garcia and Morari 1985). yiGte)— Gee =0 63) Theorem 3 Let Giz) and G{) be stable, G.(1)= 1/G(0) and Fie)= rant Oxacl G4) “Then thee exists an a* (0- 0 65) Hence, by using the structure of Fig. 13 (c) and a G,(2) such that G,(1)= 1/60),Digital controllers for SISO systems 875 stability can always be achieved for any mismatch of model and plant, as long as the steady-state gain of the model has the correct sign. This structure does not have to be actually implemented, since it is entirely equivalent to the classical feedback. The conventional feedback controller C(z) can be obtained from FUGA) T=Feexce ta Cle), ‘An advantage of using the structure of Fig. 13(c) for design purposes is that it provides a clear physical meaning to the tuning parameter a 5. Conclusions ‘The purpose of this paper was to present and explain, in a transparent manner, the problems of some well known digital controllers for SISO systems. Based on the results ofthis study, a simple rule was derived for the design of a controller which will always perform well. The reason for trying to establish a simple rule is to make sure that this design procedure is of use to the practicing engincer. ‘The algorithm developed i fee ofthe basic problems of the controllers examined, ie. intersample rippling and overshoot or undershoot. In the case where all the unstable zeros ofthe pulse transfer function ofthe system G(:) have negative real part, ityields a deadbeat controller which drives the diserete output of the system to the set- point in a finite number of time steps. When G(c) has unstable zeros with positive real part, the controller drives the output to the set-point asymptotically in order to avoid large overshoot or undershoot, When all the zeros, stable or unstable, have positive real part, it minimizes the sum of the squared errors of the output ‘An advantage of the proposed structure is that a tuning parameter is included (filter of Fig. 13 (c) whose physical meaning and effect is clear to the designer. If the ‘model is.exact, no tuning is necessary in contrast with the Dahlin and Vogel-Edgar controllers. In the case ofa mismatch between the system and the model, stability can always be guaranteed when some simple conditions hold. It is also important to rote that the same tuning parameter can be used to reduce the values of the manipulated variable at the cost of a slower response. ACKNOWLEDGEMENT Support from the National Science Foundation and the Department of Energy gratefully acknowledged. REFERENCES Astrés, K. J, HacAxpen, P, and Srenvny, J, 1984, Automatica, 2, 31. Astndxt, K. J, and WITTENwARK, B., 1980, Proc. Instn elect. Engrs, PL D, 127, 120; 1984, Computer Controlled Systems (Englewood Ciifls: Prentice-Hall Brncrn, AUR, and RaGazzint, J. Ry 1958, AV.E.E. Trans, 73, 236, CHANG, $. 5. L, 1961, Synthesis of Optimum Control Systems (New York: McGraw-Hill. Danuin, E. B, 1968, Instrum. Control Syst, 4l, 7. Garcia, C.E, and MonAkl, M,, 1982, Ind. Engng chem. Proc. Des. Dev. 21, 308; 1985, Ibid, 24, 472876 Digital controllers for SISO systems Isermawn, R, 1981, Digital Control Systems (Springer-Verlag). Ianiszowssi K., 1983, Int J. Control, 37,159. JURY, E.L, 1957, ALE. Trans, 75, 391; 1988, Sampled-Data Control Systems (New York: John Wiley) Jury, EI, and SchnoeneR, W, 1956, AEE, Trans, 75, 317. Kauwan, RE, 1984, ADEE, Trans, 73, 245, KAIMAN, RE, and BenTRas, JE. 1958, ALE.E. Trans, 78, 602 Kucera, V, 1972, Kybornetica, 8 233 Ko, B. C. 1977, Digital Control Systems (Champaign, Ilinofs: SRI Publishing Co). Luvaex, W. L., 1973, Process Modelling, Simulation and Control for Chemical Engineers (New York: McGraw-Hill, Raoazzini, J. R., and FRANKLIN, G. F, 1958, Sampled-Dara Control Systems (New York: McGraw-Hill). Surre, C. L, 1972, Digital Computer Process Control Seranton, Pennsylvanie: Intext Educa- tional Publishers) ‘Voset, E. F, 1982, Adaptive control of chemical processes with variable dead time. Ph.D. dissertation, University of Texas, Austin,